Icelandic speech to text, on your own Mac

If you have tried automatic transcription on an Icelandic meeting, you have probably seen names turned into nonsense, word endings dropped and the English half of the conversation mangled. That is not bad luck. Icelandic is genuinely hard for speech recognition, and most tools treat it as one language among a hundred.

Why Icelandic is hard for speech recognition

A small language means little training data

Modern speech recognition learns from recorded speech paired with text. English has an enormous amount of it. Icelandic, with a few hundred thousand speakers, has far less. The authors of OpenAI’s Whisper model showed that its accuracy in a language tracks how much training audio it saw in that language (Radford et al., 2022). Languages with little data come out worse, and Icelandic is one of them.

Inflection changes almost every word

Icelandic nouns, adjectives and pronouns change form by case, number and gender, and verbs change by person and tense. “Fundur”, “fund”, “fundi” and “fundar” are all the same meeting. A model has to hear the difference between endings that are often short and unstressed, and a wrong ending turns a correct sentence into a wrong one. It also means a transcript can look almost right while being wrong in ways that matter to an Icelandic reader.

Names, places and patronymics

Meetings are full of names: Guðrún Pálsdóttir, Ari Magnússon, Egilsstaðir, Hafnarfjörður. Many of them never appear in the model’s training text, and patronymics inflect like any other word (“hjá Ara Magnússyni”). The letters þ, ð and æ have to come out right too, or search and copy-paste stop working.

Icelandic and English in the same sentence

Few Icelandic workplaces speak only Icelandic. A product meeting in Reykjavík might switch to English for a colleague, use English terms for software (“deploy”, “pull request”, “roadmap”) and switch back mid-sentence. A system told to expect one language tends to force the other into it, and the result is text that is wrong in both.

What Hearkeeper does

Hearkeeper is a Mac app that records the meetings, calls and lectures you are part of and turns them into a transcript, a summary, decisions and to-dos. For Icelandic, a few things matter.

Transcription runs on your Mac

Hearkeeper runs Whisper on the Mac itself, through whisper.cpp with Metal acceleration. Audio is never uploaded unless you connect a team server or your own API key, which matters for board meetings, client calls, interviews and anything else you would not want on someone else’s server. It also works offline, in a lecture hall or on a plane.

• Live captions appear about two to four seconds behind the speaker, so you can follow along during the call.
• The final pass runs after you stop, with Whisper large-v3 turbo and speaker separation, and replaces the live text with a cleaner transcript.
• Speed: we measured the final pass at about ten times faster than real time, so an hour of Icelandic is transcribed in under six minutes on a Mac with Apple silicon.

Icelandic is first-class

Meetings default to Icelandic. The app itself ships in English and Icelandic, and every font and text field handles þ, ð, æ, ö and the accented vowels properly, so names come out as they are spelled and search finds them.

Mixed Icelandic and English is detected

When a meeting moves between Icelandic and English, Hearkeeper detects the language automatically instead of forcing everything into one. The English lines stay English and the Icelandic lines stay Icelandic. Whisper covers 99 languages in all, so a meeting with a Polish or Danish colleague works too.

Measured, not assumed

We measure Icelandic accuracy on the public FLEURS Icelandic test set, a set of read sentences published by Google (Conneau et al., 2022) that anyone can use to check a model. A public test set means the measurement can be repeated, and we can tell whether a change actually helped.

A model tuned on Icelandic speech is on the way. We will say more when we can show how it measures on the same test.

More than a transcript

Transcripts are rarely the point. After a meeting, a language model running on your Mac writes a summary, lists the decisions and pulls out the to-dos, with no API key needed. Every line links to the second it was said, so you can check an Icelandic summary against the Icelandic original. You can ask questions during the call and later across all your meetings. Everything is kept in one SQLite file in your own folder and exports to Markdown.

Recording that everyone knows about

Hearkeeper records your microphone and your computer’s audio directly on the Mac. No bot joins the call, and it works with Zoom, Google Meet, Microsoft Teams, Slack huddles, Webex, FaceTime and meetings in the room.

That does not mean recording quietly. Every recording starts with a consent step: you read a short notice aloud, post it in the chat, or both, and a consent receipt is saved with the recording. The recording pill stays on screen, shows up in screen shares, and cannot be hidden. Our guide to recording meetings and consent covers the rules in more detail, including Iceland’s.

Other options for Icelandic transcription

Hearkeeper is not the only way to transcribe Icelandic, and the right tool depends on what you need. These are the options we know of, as their makers describe them:

• Microsoft Teams lists Icelandic (Iceland) among the spoken languages for live transcription in Teams meetings (Microsoft Support). If your meetings already happen in Teams, that may be enough.
• Fireflies.ai lists Icelandic among its supported transcription languages (Fireflies Knowledge Base). It works in the cloud.
• Tiro, an Icelandic company, offers Icelandic speech recognition, including real-time captioning and captions for video (tiro.is).

Where Hearkeeper differs is in the combination: transcription and notes made on your own Mac, no bot in the call, a consent step with a receipt, and a pill that everyone can see.

Who it is for

• Teams in Iceland that meet in Icelandic and English and want notes without uploading the audio.
• Students recording lectures in Icelandic, online or offline.
• Journalists and researchers who transcribe interviews and want the audio to stay on their own machine.
• Anyone who has given up on Icelandic transcription because the names came out wrong.

Availability

Hearkeeper is not released yet. It comes to the Mac first (macOS 14.2 or later, Apple silicon and Intel), with other platforms later. Read what the website collects on the privacy page.

If you work in Icelandic and want to try it early, join the waitlist. We send one email when it is ready for you.